def read_img(img_name, width, height, keep_img_shape=config.keep_img_shape):
img_ori = tools.read_img(img_name)
if img_ori is None:
return None, None
if not keep_img_shape:
img = cv2.resize(img_ori, (width, height))
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
else:
# add by rdc ,follow @Jiachenyin1
target_h, target_w = height, width
ori_h, ori_w, _ = img_ori.shape
scale = min(target_h / ori_h, target_w / ori_w)
nw, nh = int(scale * ori_w), int(scale * ori_h)
image_resized = cv2.resize(img_ori, (nw, nh)) ## width and height
img = np.full(shape=[target_h, target_w, 3],
fill_value=0,
dtype=np.uint8)
dh, dw = (target_h - nh) // 2, (target_w - nw) // 2
img[dh:(nh + dh), dw:(nw + dw), :] = image_resized
img_ori = img
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = img.astype(np.float32)
img = img / 255.0
# [416, 416, 3] => [1, 416, 416, 3]
img = np.expand_dims(img, 0)
return img, img_ori
def __read_img(self, img_file):
'''
read img_file
return:img(RGB)
'''
img = tools.read_img(img_file)
if img is None:
return None
return img
def read_img(img_name, width, height):
img_ori = tools.read_img(img_name)
if img_ori is None:
return None, None
img = cv2.resize(img_ori, (width, height))
# img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = img.astype(np.float32)
img = img / 255.0
# [416, 416, 3] => [1, 416, 416, 3]
img = np.expand_dims(img, 0)
return img, img_ori
def read_img(self, img_file):
'''
读取 img_file, 并 resize
return:img, RGB & float
'''
img = tools.read_img(img_file)
if img is None:
return None
img = cv2.resize(img, (self.width, self.height))
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = img.astype(np.float32)
img = img/255.0
return img
def read_img(img_name, width, height):
img_ori = tools.read_img(img_name)
if img_ori is None:
return None, None
img = cv2.resize(img_ori, (width, height))
nw, nh = None, None
show_img = img
img = img.astype(np.float32)
img = img/255.0
# [416, 416, 3] => [1, 416, 416, 3]
img = np.expand_dims(img, 0)
return img, nw, nh, img_ori, show_img
def read_img(self, img_file):
'''
读取 img_file, 并 resize
return:img, RGB & float
'''
img = tools.read_img(img_file)
if img is None:
return None
if self.keep_img_shape:
# keep image shape when we reshape the image
img, new_w, new_h = data_augment.keep_image_shape_resize(
img, size=[self.width, self.height])
else:
img = cv2.resize(img, (self.width, self.height))
new_w, new_h = None, None
if self.flip_img:
# flip image
img = data_augment.flip_img(img)
if self.gray_img and (np.random.random() < 0.2):
# probility of gray image is 0.2
img = data_augment.gray_img(img)
if self.erase_img and (np.random.random() < 0.3):
# probility of random erase image is 0.3
img = data_augment.erase_img(img, size_area=[20, 100])
if self.invert_img and (np.random.random() < 0.1):
# probility of invert image is 0.1
img = data_augment.invert_img(img)
if self.rotate_img:
# rotation image
img = data_augment.random_rotate_img(img)
test_img = img
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = img.astype(np.float32)
img = img / 255.0
return img, new_w, new_h, test_img
def read_img(img_name, width, height):
img_ori = tools.read_img(img_name)
if img_ori is None:
return None, None
if config.keep_img_shape:
img, nw, nh = data_augment.keep_image_shape_resize(
img_ori, size=[width, height])
else:
img = cv2.resize(img_ori, (width, height))
nw, nh = None, None
show_img = img
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = img.astype(np.float32)
img = img / 255.0
# [416, 416, 3] => [1, 416, 416, 3]
img = np.expand_dims(img, 0)
return img, nw, nh, img_ori, show_img
def read_img(self, img_file):
'''
read img_file, and resize it
return:img, RGB & float
'''
img = tools.read_img(img_file)
if img is None:
return None, None, None, None
ori_h, ori_w, _ = img.shape
img = cv2.resize(img, (self.width, self.height))
# flip image
if np.random.random() < self.flip_img:
self.is_flip = True
img = cv2.flip(img, 1)
# gray
if np.random.random() < self.gray_img:
tmp = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# convert to 3 channel
img = cv2.cvtColor(tmp, cv2.COLOR_GRAY2BGR)
# random erase
if np.random.random() < self.erase_img:
erase_w = random.randint(20, 100)
erase_h = random.randint(20, 100)
x = random.randint(0, self.width - erase_w)
y = random.randint(0, self.height - erase_h)
value = random.randint(0, 255)
img[y:y + erase_h, x:x + erase_w, :] = value
test_img = img
img = img.astype(np.float32)
img = img / 255.0
# gasuss noise
if np.random.random() < self.gasuss:
noise = np.random.normal(0, 0.01, img.shape)
img = img + noise
img = np.clip(img, 0, 1.0)
return img, ori_w, ori_h, test_img
